One piece at a time on what strategy research tends to gloss over. Free to read; PDFs linked below.
A cross-sectional long-short factor momentum portfolio built on 153 U.S. characteristic factors earns nearly all of its 1970–2024 cumulative log return inside five non-overlapping thirty-six-month windows, each followed by a twelve-month reversal of the same sign. The timing is regular; the compositions are not. Pairwise Jaccard overlap of the top-ten contributors is bounded by 0.25, and the rolling principal subspace of the factor correlation matrix at these episodes is statistically indistinguishable from quiet periods — so a risk lens trained on factor identities or covariance regimes is unlikely to see the crashes coming.
Factor backtests treat the investor as a price-taker. Recast as a noisy informed trader in a multi-investor Kyle (1985) auction, the share of gross alpha lost to price impact takes a closed form — an erosion ratio that depends only on signal informativeness and the number of competing investors. Fit to ten-year smoothed returns of eight canonical U.S. equity factors over 1950–2025, the same expression sorts them cleanly into tiers; out-of-sample, calibrating only on pre-2010 data, it beats both the historical mean and a linear time-trend on six of eight factors.